Optimal Asset Allocation in Asset Liability Management

Transcription

1 Optimal Asset Allocation in Asset Liability Management Jules H. van Binsbergen Stanford GSB and NBER Michael W. Brandt Fuqua School of Business Duke University and NBER This version: June 2012 Abstract We examine the impact of regulations on the investment decisions of a defined benefits pension plan. We show that financial reporting rules have real effects on investment behavior. For example, the current requirement to discount liabilities at a rolling average of yields, as opposed to at current yields, encourages excessive risk taking and increases the demand for short-term debt. We further assess the influence of ex ante (preventive) and ex post (punitive) risk constraints on the gains to dynamic, as opposed to myopic, decision making. We find that preventive measures, such as Value-at-Risk constraints, tend to decrease the gains to dynamic investment. In contrast, punitive constraints, such as mandatory additional contributions from the sponsor when the plan becomes underfunded, lead to very large utility gains from solving the dynamic program. We thank Tim Bollerslev, Frank de Jong, Joachim Inkmann, Ralph Koijen, Vinay Nair, Theo Nijman, Anamaría Pieschacón, George Tauchen, Bas Werker, and seminar participants at the 2005 SAMSI conference on Financial Mathematics, Statistics and Econometrics, Duke University, Tilburg University and ABP Investments for helpful discussions and comments. Jules van Binsbergen thanks the Prins Bernhard Cultuurfonds for generous financial support. Stanford, CA Phone: (650) jvb2@gsb.stanford.edu. Durham, NC Phone: (919) mbrandt@duke.edu.

2 1 Introduction We examine the impact of existing regulations on the investment decisions of a defined benefits pension plan. A recent study by the Center of Retirement Research estimates that by the end of 2008, the aggregate funding status of defined benefit plans has fallen to 75% of total liabilities and that more than 50% of private defined benefit plans are less than 80% funded. 1 Recently a set of funding rule reforms has been proposed to strengthen the pension system. These proposed reforms include limits on the deviation of the actuarial values of assets and liabilities from their market values and more stringent guidelines for pension plans to make up shortfalls in the value of their assets, relative to that of their liabilities, through additional financial contributions (AFCs) to the plan by the plan sponsor. We study in this paper the optimal asset allocation decisions of an investment manager of a defined benefit pension plan as a function of the plan s funding ratio (defined as the ratio of its assets to liabilities), interest rates, and the equity risk premium. We compare the optimal investment decisions under several policy alternatives to understand better the real effects of financial reporting and risk management rules. We focus on two mechanisms to keep investment managers from taking undesirable actions: prevention and punishment. Ex ante (preventive) risk constraints, such as Value-at- Risk (VaR) constraints, short-sale constraints and a maximum weight in stocks, restrict the investment manager s set of allowable portfolio weights. The manager is required to adhere to these ex ante constraints but is not held responsible for bad return realizations ex post. In contrast, ex post (punitive) risk constraints, when triggered, lead to a punitive action that decreases the investment manager s utility either through loss of personal compensation or reputation. Such ex post constraints do not restrict investment choices ex ante, but the manager is held accountable for a bad return realizations ex post. The ex post constraint we focus on, is the requirement to draw AFCs from the plan sponsor whenever the plan becomes underfunded. 2 At first glance, ex ante and ex post risk constraints may seem similar as both aim to decrease the risk-taking behavior of the manager. However, we show that they have profoundly different implications for the gains to dynamic, as opposed to myopic, decision making. We show that ex ante (preventive) constraints tend to decrease the gains to dynamic investment. Ex post (punitive) constraints, in contrast, largely increase the utility gains from solving the dynamic program. In other words, under ex ante constraints, the myopic solution 1 See Munnell, Aubry and Muldoon (2008) 2 Another important example of an ex post constraint is a firing rule for investment managers based on their performance.

3 provides a good approximation for the optimal solution whereas under ex post constraints it requires dynamic optimization to make the optimal investment decision. As such, ex post constraints induce the manager to behave strategically. Another important aspect of the asset liability management (ALM) problem is the discount factor used for computing the present value of a pension plan s liabilities. Recently, the discount factor has received a lot of attention for its impact on the reported financial position of the plan. On one hand, discounting by current yields guarantees an accurate description of the fund s financial situation. On the other hand, using a constant yield smoothes out temporary fluctuations in the present value of the liabilities and gives a more long-term description of the fund s financial condition. Under current regulations, the discount factor equals a four-year rolling average of the 30-year government (or corporate) bond yield, which constitutes a compromise between the two options described above. What has received much less attention, however, is the effect that these financial reporting rules can have on the optimal decisions of the investment manager of the pension plan. show that the way liabilities are computed can drive an important wedge between the fund manager s long-term objective of maximizing the funding ratio and his short-term objective (and/or requirement) of satisfying risk constraints and avoiding AFCs from the plan sponsor as described above. The key to this wedge is the fact that these risk constraints are based on the (smoothed) reported liabilities instead of on the actual liabilities that (should) enter into the investment manager s long-term objective. We thus examine two important issues in a stylized ALM problem. First, we address the role of hedging demands. We ALM problems are inherently long-horizon problems with potentially important strategic aspects. 3 They differ from standard portfolio choice problems (Markowitz (1952), Merton (1969,1971), Samuelson (1969) and Fama (1970)), not only because of the short position in the pension liabilities, but also because of the regulatory risk constraints and mandatory AFCs discussed above. We assume that the investment manager dislikes drawing AFCs from the plan sponsor and directly model this dislike as a utility cost. We interpret this utility cost as a reduced form for the loss of compensation or reputation of the investment manager. In other words, drawing mandatory AFCs serves as an ex post (punitive) risk constraint. The associated utility cost introduces a kink in the value function of the investment manager s dynamic optimization problem that causes the manager to become first-order risk averse whenever the (reported) funding ratio approaches the critical threshold that triggers AFCs. We show that this kink in the value function leads to substantial hedging demands and large certainty equivalent utility gains from dynamic 3 Recent strategic asset allocation studies include Kim and Omberg (1996), Campbell and Viceira (1999), Brandt (1999,2005), Aït-Sahalia and Brandt (2001), and Sangvinatsos and Wachter (2005).

4 investment. We also find that ex ante risk constraints, such as Value-at-Risk constraints, decrease the gains to dynamic investment. Such constraints do not introduce a kink in the value function, leading to the relatively flat peak of the value function that is common for intertemporal investment problems with power utility. This flat peak implies that even large deviations from the optimal decision lead to only small welfare losses (e.g., Cochrane (1989)). In addition, the ex ante constraints decrease the menu of available portfolio weights, thereby inhibiting the manager to execute his desired hedging demands. As a consequence the gains from dynamic (strategic) investment are small. Second, we show that under the assumption that the plan manager dislikes AFCs and in the long run wishes to maximize (the utility from) the actual funding ratio, smoothing the reported liabilities induces grossly sub-optimal investment behavior. The associated welfare loss is a direct consequence of the misalignment of long- and short-term incentives. The investment manager makes a tradeoff between his long-term objective of maximizing the actual funding ratio (assets to actual liabilities) and his short-term concerns about risk constraints and drawing AFCs based on the reported funding ratio (assets to smoothed liabilities). Interestingly, we find that risk controls based on the smoothed liability measure can inadvertently induce the manager to take riskier, instead of less risky, positions. The investment behavior of corporate pension plans has been studied by Sundaresan and Zapatero (1997) and by Boulier, Trussant and Florens (2005). Sundaresan and Zapatero (1997) model the marginal productivity of the workers of a firm and solve the investment problem of its pension plan assuming a constant investment opportunity set consisting of a risky and a riskless asset. We instead allow for a time-varying investment opportunity set including cash, bonds, and stocks. More importantly, we consider the ALM problem from the perspective of the investment manager as a decision maker and investigate how regulatory rules influence the optimal investment decisions. In order to focus attention on the asset allocation side of the ALM problem, we model the liabilities of the pension plan in reduced form by assuming a constant duration of 15 years. Boulier, Trussant, and Florens (2005) also assume a constant investment opportunity set with a risky and a riskfree asset. In their problem, the investment manager chooses his portfolio weights to minimize the expected discounted value of the contributions over a fixed time horizon, with the constraint that the value of the assets cannot fall below that of the liabilities at the terminal date. This problem setup implicitly assumes that the pension plan terminates at some known future date and that the investment manager s horizon is equal to this terminal date. By taking the investment manager s preferences and horizon as the primitive, our perspective is different. The manager has a motive to minimize (the

5 disutility from) the sponsor s contributions, captured by the AFCs in our case. However, the manager also has a bequest motive by wanting to maximize the funding ratio at the end of his investment horizon. The end of the manager s investment horizon may be long before the pension plan terminates, which is why we hold the duration of the liabilities fixed. Our contributions to the literature are the following. First, we attempt to bridge further the gap between the dynamic portfolio choice literature and the ALM literature. 4 We pose the ALM problem as a standard dynamic portfolio choice problem by defining terminal utility over the ratio of assets and liabilities, as opposed to over assets only. This approach allows a parsimonious representation of the ALM problem under a time-varying investment opportunity set. Solving this dynamic program is relatively straightforward compared to the usual, more complicated, stochastic programming techniques. We then assess the interplay between dynamic hedging demands, risk constraints, and first-order risk aversion. show that the solution to the ALM problem under ex post (punitive) constraints involves economically significant hedging demands, whereas ex ante (preventive) constraints decrease the gains from dynamic investment. Finally, we explicitly model the trade-off between the long-term objective of maximizing terminal utility and the short-term objective of satisfying VaR constraints and avoiding AFCs from the plan sponsor. 5 We We show that if these shortterm objectives are based on reported liabilities that are different from actual liabilities, they can lead to large utility losses with respect to the long-term objective. In order to focus on these contributions, there are at least three important aspects of the ALM problem that we do not address explicitly. First, there is a literature, starting with Sharpe (1976), that explores the value of the so-called pension put arising from the fact that U.S. defined-benefit pension plans are insured through the Pension Benefit Guarantee Corporation. Sharpe (1976) shows that if insurance premiums are not set correctly, the optimal investment policy of the pension plan may be to maximize the difference between the value of the insurance and its cost. This obviously induces perverse incentives. Second, we do not incorporate inflation. Besides affecting the allocation to real versus nominal assets (Hoevenaars et al. (2004)), inflation drives another wedge between the long-term objective of maximizing the real funding ratio, computed with liabilities that are usually pegged to real wage levels, and the short-term objective of satisfying risk controls and avoiding AFCs 4 Campbell and Viceira (2002) and Brandt (2005) survey the dynamic portfolio choice literature. 5 We could easily incorporate other short-term objectives, such as beating a benchmark portfolio over the course of the year (see also Basak, Shapiro, and Teplá (2006) and Basak, Pavlova, and Shapiro (2007)). Whenever this short-term objective is defined with respect to reported liabilities that are different from actual liabilities, this leads to a similar misalignment of incentives as the one we explore in this paper. It is interesting to note that in practice pension fund managers are often assessed relative to an assets-only benchmark, which is a benchmark that implicitly assumes constant liabilities (see van Binsbergen, Brandt and Koijen (2008)).

6 based on nominal valuations. Third, we ignore the taxation issues described by Black (1980) and Tepper (1981). The paper proceeds as follows. Section 2 describes the return dynamics, the preferences of the investment manager, and the constraints under which the manager operates. We model the dynamics of stock returns, short-term bond yields, and long-term bond yields, as a first-order vector autoregression (VAR). The investment manager has power utility with constant relative risk aversion (CRRA) but incurs a linear utility cost every time the pension plan becomes underfunded and is forced to draw AFCs from its sponsor. Section 3 describes our numerical solution method for the dynamic optimization problem, which is a version of the simulation-based algorithm developed by Brandt, Goyal, Santa-Clara, and Stroud (2005). Section 4 presents our results. Because the impact of yield smoothing can easily be illustrated in a simple one-period model, we first address this issue before addressing the role of hedging demands in the multi-period setting. We show that smoothing yields to compute the value of the liabilities can lead to grossly suboptimal investment decisions. We then assess the gains to dynamic, as opposed to myopic investment for four cases: i. a standard CRRA investment problem with a time-varying investment opportunity set (no liabilities, no sponsor contributions, no risk constraint), ii. a standard ALM problem (no sponsor contributions and no risk constraint), iii. an ALM problem with a VaR constraint (no sponsor contributions), and iv. an ALM problem with sponsor contributions. We show that ex ante risk constraints decrease the already small gains from dynamic investment in the absence of AFCs. However, when we introduce the utility cost of AFCs, the gains from dynamic investment become economically very large. 2 ALM problem The ALM problem requires that we specify the investment opportunity set (or return dynamics), the preferences of the investment manager, and the risk constraints the investment manager faces. The next three sections describe these three items in turn. 2.1 Stochastic Discount Factor and Bond Returns We consider a pension plan that can invest in three asset classes: stocks, bonds, and the riskfree asset. Stocks are represented by the Standard and Poors (S&P) 500 index, bonds by a 15-year constant maturity Treasury bond, and the riskfree asset by a one-year Treasury bill. We consider an annual rebalancing frequency. We reduce the investment opportunity set to three asset classes driven by two state variables, to keep the dimensionality of the

7 problem low. Considering only three asset classes may seem restrictive, however, these asset classes should be interpreted as broader categories where long-term bonds represent assets that are highly correlated with the liabilities; stocks and one-year Treasury bills represent assets that have a low correlation with liabilities and have respectively a high risk/high return and low risk/low return profile. We assume that the one-year and 15-year yield levels follow a first-order VAR process. We model stock returns with a time-varying risk premium that depends on the level and slope of the yield curve (e.g., Ang and Bekaert (2005)). We model the return dynamics as follows. Following Ang and Piazzesi (2003), let the stochastic discount factor be given by: m t+1 = exp( y (1) t 1 2 ΛT t Λ t Λ T t ε t+1 ) (1) where Λ t are the market prices of risk for the various shocks, which we assume are a linear function of a vector of state variables X t : Λ t = Λ 0 + Λ 1 X t (2) where the demeaned state variables follow a VAR(1) model: X t = ΦX t 1 + Σε t (3) with ε t N (0, I), such that the variance of the shocks is given by ΣΣ T. As the state variables we choose the (demeaned) 1-year and 15-year yields: X t = [ X 1,t X 2,t ] = [ y (1) t µ y,1 y (15) t µ y,15 ] (4) The mean vector µ is given by: µ = [ µ y,1 µ y,15 ] The dimension of the parameter vectors of the prices of risk are as follows: Λ 0 is a 2x1 vector, Λ 1 a 2x2 matrix and ΣΣ T a 2x2 matrix. Let p (n) t denote the zero coupon bond price with maturity n at time t. The Euler condition for bond prices is then given by: [ ] p (n) t = E t m t+1 p (n 1) t+1 (5)

8 Recursively solving this relationship results in bond prices that are exponential linear functions of the state vector X t : p (n) t = exp (A n + B n X t ) (6) where the coefficients follow the recursion: A n+1 = A n B T n ΣΛ BT n ΣΣ T B n B n+1 = (Φ ΣΛ 1 ) B n e 1 [ ] T with e 1 = 1 0 The initial conditions are given by A1 = µ y,1 and B 1 = e 1. Further, the fact that the second state variable is itself a yield, leads to the following restrictions: A 15 = µ y,15 and B 15 = e 2. Bond yields are then given by: y (n) t = log ( ) p (n) t = a n b n X t (7) n 2.2 Stochastic Discount Factor and Stock Returns We model the dynamics for log stock returns as follows: r s,t+1 = µ s + βx t + η t+1 with η t+1 N ( ) 0, ση 2 (8) where the shock η t+1 is correlated with the two shocks in ε t+1. Let ρ i denote the correlation between η t+1 and the i th element in ε t+1. Recall that the state vector X t is demeaned, implying that µ s is the average return on stocks. As, under no arbitrage, stocks are priced by the pricing kernel, this leads to the following additional restrictions: E t (m t+1 exp(r s,t+1 )) = 1. (9) Due to the exponential setup, this results in the following three restrictions after substituting 8 and 1 into 9: [ ] µ s µ y, σ2 η = Λ T ρ 1 σ η 0 ρ 2 σ η [ ] β e T 1 = Λ T ρ 1 σ η 1 ρ 2 σ η

9 The first equation states that the log equity risk premium µ s µ y, σ2 η is equal to the unconditional covariance between unexpected stock returns η t+1, and the shocks to the stochastic discount factor. The second equation links the time variation in the equity risk premium to the time variation in the two state variables. 2.3 Estimation The return dynamics we propose allow for both a time-varying risk free rate, time-varying expected bond returns, and a time-varying equity risk premium, all as a function of two state variables, the short-term and the long-term yield. 6 The parameters of the model Θ are given by: Θ = {Λ 0, Λ 1, µ, Φ, Σ, α, β, σ η, ρ 1, ρ 2 } which is a total of 21 parameters. The parameters can be estimated in one stage using OLS by estimating the following system: r s,t y (1) t y (15) t [ = Γ 0 + Γ 1 y (1) t 1 y (15) t 1 ] + ξ t with ξ t MVN (0, Θ), (10) From the OLS estimates ˆΓ 0, ˆΓ 1 and the estimated covariance matrix ˆΘ, it is straightforward to uncover µ, Φ, α, β, Σ, ρ 1, ρ 2 and σ η. The remaining parameters that need to be determined are the 6 parameters in Λ 0 and Λ 1. The 6 parameters in Λ 0 and Λ 1 are implied by 6 equations. The first three equations follow from the fact that the 15-year yield is a state variable: A 15 = µ y,15 and B 15 = e 2. The remaining three equations follow from the Euler equation for stocks given in (10). The estimates are presented in Appendix A. 2.4 Liabilities and Simple Returns We assume that the pension plan has liabilities with a fixed duration of 15 years. We measure the value of these liabilities in three ways. First, we compute the actual present value of the 6 For recent work on return predictability see Binsbergen and Koijen (2010), Ang and Bekaert (2005), Lewellen (2004), Campbell and Yogo (2005), and Torous, Valkanov, and Yan (2005) for stock returns, as well as Dai and Singleton (2002) and Cochrane and Piazzesi (2005) for bond returns.

10 liabilities by discounting by the actual 15-year government bond yield: 7 L t = exp( 15y 15,t ). (11) Our second measure is based on current regulations prescribing that the appropriate discount factor is the four-year average long-term bond yield, which implies: where ˆL t = exp( 15ŷ 15,t ), (12) ŷ 15,t = 1 4 (y 15,t + y 15,t 1 + y 15,t 2 + y 15,t 3 ). (13) Finally, we compute the value of the liabilities using a constant yield equal to the steady state value of the long-term bond yield ȳ 15 implied by the VAR (see Appendix A): L t = exp( 15ȳ 15 ) (14) Note that with all three measures the liabilities follow a stationary stochastic process. The model could easily be extended to include a deterministic time trend representing demographic factors. However, to maintain a parsimonious representation, we focus on the detrended series. Our specification also abstracts from inflows (premium payments) and outflows (pension payments) to the fund. We assume that in each year the inflows equal the outflows, which allows us to focus purely on the investment management part of the fund. The only inflows we consider are cash injections by the plan sponsor required to meet the regulator s minimal funding level. Note further that the three measures of liabilities described above are driven by only one risk factor, the 15-year government bond yield. This could suggest that a one-factor model for the term structure would suffice in our model. However, we assume a two-factor model to allow for a time-varying riskfree rate. We compute the simple gross returns on the three asset classes as follows: R f,t = exp(y 1,t 1 ) (15) [ ] [ ] R s,t exp(r s,t ) R t = =, exp( 14y 14,t )/ exp( 15y 15,t 1 ) R b,t where R s,t is the simple gross return on stocks, R f,t is the return on the one-year T-bill 7 To maintain a parsimonious representation we use the 15-year bond yield to determine the discount factor instead of the 30-year bond yield. Since the dynamics of both yields are very similar, this simplification does not influence our results.

11 (riskfree), and R b,t is the simple gross return on long-term bonds. The funding ratio of the pension plan is defined as the ratio of its assets to liabilities: where assets evolve from one period to the next according to: S t = A t L t, (16) A t = A t 1 (R f,t + α t 1 (R t R f,t )) + c t exp( 15y 15,t ) for t 1 (17) and α t [α s,t, α b,t ] denotes the portfolio weights in stocks and bonds. We let c t denote the contributions of the plan sponsor at time t as a percentage of the liabilities which, under actual discounting, are equal to exp( 15y 15,t ). Note that defining the contributions as a percentage of the liabilities is equivalent to expressing contributions in future (t + 15) dollars. When liabilities are determined through constant discounting or four-year average discounting, we define the contributions as a percentage of those liability measures and the last term in expression (17) is adapted accordingly. We use Ŝt and S t to denote the funding ratios computed using the liability measures ˆL t and L t, respectively. S t Finally, we define A t as the assets in period t before the contributions are received, and as the ratio of A t and the liabilities: 2.5 Preferences A t = A t 1 (R f,t + α t 1 (R t R f,t )) for t 1 (18) S t = A t L t (19) We take the perspective of an investment manager facing a realistic regulatory environment. We assume that the manager s utility is an additively separable function of the funding ratio at the end of the investment horizon (S T ) and the requested extra contributions from its sponsor as a percentage of the liabilities (c t ). We assume that the manager suffers disutility in the form of unmodeled reputation loss or loss in personal compensation for requesting these contributions. The utility function of the manager is given by: ( ) U S T, {c t } T 1 t=1 =E 0 [u (S T ) =E 0 [ β T S1 γ T ] T v (c t, t) t=1 ] T 1 γ λ β t c t where γ 0 and λ 0. t=1 (20)

12 The first term in the utility function is the standard power utility specification with respect to the funding ratio at the end of the investment horizon, S T. We call this wealth utility. We assume that this wealth utility always depends on the actual funding ratio. That is, we use the actual yields to compute the liabilities in the denominator, regardless of government regulations, as opposed to using a smoothed or constant yield. The motivation for this assumption is that ultimately the manager is interested in maximizing the actual financial position of the fund, which is also the position the pension holders care about. 8 The term T t=1 βt v (c t ) represents the investment manager s disutility (penalty) for requesting and receiving extra contributions c t from the plan sponsor. This penalty can be interpreted as loss in reputation or compensation. The linear function just reflects the first-order effect of these penalties and higher order terms could be included in our analysis. 9 Furthermore, when contributions from the sponsor are set equal to the funding ratio shortfall, linearity of the function v( ) implies that the utility penalties are scaled versions of the expected loss, which, next to a VaR constraint, is often used as a risk measure. As noted by Campbell and Viceira (2005), the weakness of a VaR constraint is that it treats all shortfalls greater than the VaR as equivalent, whereas it seems likely that the cost of a shortfall is increasing in the size of the shortfall. They, therefore, propose to incorporate the expected loss directly in the utility function, which in our framework is achieved by the linearity of the function v( ). Finally, the investment manager discounts next period s utility and disutility by the subjective discount factor β. Another appealing interpretation of our utility specification is the following. In the context of private pension plans, the investment manager acts in the best interest of two stakeholders of the plan, (i) the pension holders who are generally risk averse and (ii) the sponsoring firm which we assume to be risk neutral. The parameter λ then measures the investment manager s tradeoff between these two stakeholders. If one believes that the investment manager merely acts in the best interest of the firm, the value of λ is high. Conversely, if one believes that the investment manager acts mainly in the interest of the 8 It is interesting to note that even when both wealth utility and the risk constraints/afcs are determined through four-year average discounting, there is still a misalignment of incentives for a multiperiod investment problem. The risk constraints (which apply in every period) still induce the use of the risk-free asset. This is a consequence of the large reduction of the conditional variance that yield smoothing induces: var t (ŷ 15,t+1 ) = var t [ 1 4 (y 15,t+1 + y 15,t + y 15,t 1 + y 15,t 2 )] = 1 16 var t[y 15,t+1 ]. Wealth utility, on the other hand, depends on the funding ratio in year T. The conditional variance of ŷ 15,T is given by: var t (ŷ 15,T ) = var t [ 1 4 (y 15,T + y 15,T 1 + y 15,T 2 + y 15,T 3 )]. Note that for a 10-year investment problem, T = 10, the yields in year 10, nine, eight and seven (which jointly determine the liabilities in year 10) are all unknown before year seven. Therefore, in periods one through six, long-term bonds are still the preferred instrument to hedge against liability risk when maximizing wealth utility. 9 Non-linear specifications for the function v( ), e.g. a quadratic form, do not change our qualitative results.

13 beneficiaries, λ is low. Finally, we can interpret the proposed utility specification in yet two other interesting ways. First we can interpret it as a portfolio choice problem with intermediate consumption and bequest. In the literature on life-time savings and consumption, it is common to assume that utility from consumption is additively separable from bequest utility. The only difference is that, in our case, consumption is strictly negative and not strictly positive. In other words, the investment manager can increase his wealth by suffering negative consumption which leads to a tradeoff between maximizing (the utility from) the funding ratio at the end of the investment horizon and minimizing (the disutility from) the contributions along the way. The second interpretation is that similar utility specifications have been used in the general equilibrium literature with endogenous default, where agents may choose to default on their promises, even if their endowments are sufficient to meet the required payments (e.g., Geanakoplos, Dubey, Shubik (2005)). Agents incur utility penalties which are linearly increasing in the amount of real default. The idea of including default penalties in the utility specification was first introduced by Shubik and Wilson (1977). The tradeoff between the disutility from contributions and wealth utility is captured by the coefficient λ. When we impose that in each period the sponsor contributions are equal to the funding ratio shortfall, and this shortfall is determined through actual discounting, a value of λ = 0 implies that the investment manager owns a put option on the funding ratio with exercise level S = 1. This gives the manager an incentive to take riskier investment positions. When λ, the disutility from contributions is so high that the investment manager will invest conservatively to avoid a funding ratio shortfall when the current funding level is high. Depending on how liabilities are computed, investing conservatively either implies investing fully in the riskfree asset or investing fully in bonds (to immunize the liabilities) or a mixture of the two. 10 Increasing the funding ratio at time zero affects the expected utility in three ways. First, it increases current wealth and therefore, keeping the investment strategy constant, also increases expected wealth utility. Second, if there is a period-by-period risk constraint, a higher funding ratio will make the risk constraint less binding in the current period and also decreases its expected impact on future decisions. Third, keeping the investment strategy constant, the probability of incurring contribution penalties in future periods decreases. 10 When λ 1, concavity of the utility function is guaranteed under actual discounting. For λ = 1, the utility is smooth, but for λ > 1, it is kinked at S = 1. The right derivative of the function 1 1 γ [max(s, 1)] 1 γ λmax(1 S, 0) is 1 whereas its left derivative equals λ. The risk neutrality over losses combined with the kinked utility function at S = 1 resembles elements of prospect theory (Kahneman and Tversky (1979)).

14 2.6 Constraints Short sale constraints We assume that the investment manager faces short sales constraints on all three assets: α t 0 and α tι 1. (21) VaR constraints Pension funds often operate under Value-at-Risk (VaR) constraints. A VaR constraint is an ex ante (preventive) risk constraint. It is a risk measure based on the probability of loss over a specific time horizon. For pension plans, regulators typically require that over a specific time horizon the probability of underperforming a benchmark is smaller than some specified probability. The most natural candidate for this benchmark is the fund s liabilities. In this case, the VaR constraint requires that in each period the probability of being underfunded in the next period is smaller than probability δ. We set δ equal to Depending on prevailing regulations, the relevant benchmark can be the actual liabilities (L t ), constant liabilities ( L t ) or, as under current regulations, ˆL t. We also compute the optimal portfolio weights and certainty equivalents when there are no additional contributions from the sponsor. In that case, there is no external source of funding that guarantees the lower bound of one on the funding ratio. It may therefore be that in some periods the fund is underfunded to begin with. In those cases, the VaR constraint described above can not be applied and requires adaptation. When at the beginning of the period the fund has less assets than liabilities, we impose that the probability of a decrease in the funding ratio is less than In other words, if the fund is underfunded to begin with, the manager faces a VaR constraint as if the funding ratio equals one AFCs Under current regulations, a pension plan is required to receive AFCs from its sponsor whenever it is underfunded. As the manager dislikes drawing AFCs from the plan sponsor, this requirement serves as an ex post (punitive) risk constraint. The government regulation around these mandatory AFCs is not at all trivial. First, the fund is allowed to amortize a realized shortfall over 30 years. Current reform proposals shorten this amortization to 18 years. Furthermore, there is currently a credits system in place which implies that previous excess contributions can be subtracted from current required contributions, regardless of the

15 financial condition of the fund. For example, if a fund is 30 percent underfunded, but in the past the plan sponsor has contributed significantly more than necessary, current regulations exempt the sponsor from having to reduce the shortfall. This credits system can obviously significantly endanger the financial stability of the pension system, which is by now well recognized and major changes have been proposed. In our setup, we set the contributions of the sponsor equal to the funding ratio shortfall in each period. However, the measurement of this shortfall is highly dependent on the way liabilities are computed, which is what we study in this paper. Hence, we do not allow for credits nor do we allow for amortization of the shortfall. Since the latter can easily be mimicked by a bond that amortizes over time, we do not consider this to be a severe restriction in our model. 2.7 Data description and estimation We use annual data from 1954 through 2010 to estimate the parameters of the return process. For stock returns we take the natural logarithm of the return on the S&P 500 composite index including distributions. For bond yields we use the continuously compounded constant maturity yields as published by the Federal Reserve Bank. Whenever data on 15-year government bonds is missing, we take an average of the 10 and 20-year bond yields. We estimate the model by OLS. 11. The estimation results are given in Appendix A. In Figure 1 we plot: i. the 15-year bond yield, ii. the four-year smoothed 15-year bond yield and, iii. the steady state value for the 15-year bond yield that follows from our VAR specification. The graph shows that the unconditional variance of the 15-year bond yield is close to the unconditional variance of the smoothed 15-year bond yield. In other words, the 15-year yield is so persistent that a four-year smoothing period is not long enough to decrease its unconditional variance. 12 To the extent that the purpose of yield smoothing is to create stability in the pension system by decreasing the unconditional variance of the discount factor, we have to conclude that this goal is currently not reached. However, the conditional variance of the smoothed series, given by: var t (ŷ 15,t+1 ) = var t [ 1 4 (y 15,t+1 + y 15,t + y 15,t 1 + y 15,t 2 )] = 1 16 var t[y 15,t+1 ], (22) is a factor sixteen smaller than the conditional variance of the actual yield series. conditional variance reduction in combination with the risk constraints induces the perverse incentives that are the scope of this paper. 11 Include dummy variables for the period in our estimation to correct for this exceptional period with high inflation does not change our findings or conclusions 12 As noted before, similar results hold for the 30-year bond yield This

16 3 Method The ALM investment problem, even in stylized form, is a complicated and path-dependent dynamic optimization program. We use the simulation-based method developed by Brandt, Goyal, Santa-Clara, and Stroud (2005) to solve this program. The main idea of their method is to parameterize the conditional expectations used in the backward recursion of the dynamic problem by regressing the stochastic variables of interest across simulated sample paths on a polynomial basis of the state variables. 13 More specifically, we generate N = 10, 000 paths of length T from the estimated return dynamics. We then solve the dynamic problem recursively backward, starting with the optimization problem at time T 1: [ ] β max U (S T ) = max E T 1 α T 1 α T 1 1 γ S1 γ T λβc T, (23) subject to equations (10), (11), (15), (16) and (17) as well as the short sale constraints and the definition of the required contributions. The solution of this problem depends on S T 1. To recover this dependence, we solve a range of problems for S T 1 varying between 0.4 and three. For each value of S T 1 we optimize over the portfolio weights α T 1 by a grid search over the space [0, 1] [0, 1]. This grid search over the portfolio weights avoids a number of numerical problems that can occur when taking first order conditions and iterating to a solution. We then evaluate the conditional ( ) expectation E T 1 S 1 γ T by regressing for each value of ST 1 and each grid point of α T 1 the realizations of S 1 γ T (N 1) on a polynomial basis of the two state variables y 15,T 1 and y 1,T 1. Define: [ ] [ ] z = z 1 z 2 = y 1,T 1 y 15,T 1, (24) then and where X = 1 z 1,1 z 2,1 (z 1,1 ) 2 (z 2,1 ) 2 (z 1,1 ) (z 2,1 )... 1 z 1,2 z 2,2 (z 1,2 ) 2 (z 2,2 ) 2 (z 1,2 ) (z 2,2 ) z 1,N z 2,N (z 1,N ) 2 (z 2,N ) 2 (z 1,2 ) (z 2,N )... (25) ( ) E T 1 S 1 γ T = X ˆβ, (26) ˆβ = (X X) 1 X ( ) S 1 γ T. (27) 13 This approach is inspired by Longstaff and Schwartz (2001) who first proposed this method to price American-style options by simulation.

17 When liabilities are discounted using the rolling four-year average yield, we have to include polynomial expansions of all four lags of both state variables in our solution method. To evaluate the conditional expectation of the contributions in period T, E T 1 (c T ), we first regress 1 S T on X: ˆζ = (X X) 1 X (1 S T ). (28) Assuming normality for the error term in the regression and letting ˆσ denote its standard deviation, we find: E T 1 (c T ) = E T 1 [max (1 S T, 0)] = Φ ( ) ( ) X ˆζ X X ˆζ ˆζ + ˆσφ, (29) ˆσ ˆσ where Φ ( ) and φ ( ) denote respectively the cumulative and probability density functions of the standard normal distribution. Note that this conditional expectation also represents the expected loss of the fund over the next period. In the exposition above, contributions are determined by actual discounting. When we use constant or four-year average discounting we simply replace S T by S T or ŜT and we replace S T by S T or Ŝ T. Given the solution at time T 1, meaning the mapping from S T 1 to the optimal α T 1, we iterate backwards through time. The iterative steps are as described above with just a few additions. For ease of exposition we now describe these additions for period T 2, but they equally apply for periods T 3, T 4,..., 1. At time T 2 we determine for each grid point of α T 2 the return on the portfolio in path i N from T 2 to T 1. Using this return to compute S T 1,i, we can then compute the return in path i from T 1 to T by interpolating over the mapping from S T 1 to α T 1 derived in the previous step. Similarly, we interpolate in each path the expected penalty payments. We impose in each path and in each period a VaR constraint. For given values of S t we determine for each α t the conditional mean and conditional variance of the funding ratio in period t + 1 through regressions on the polynomial basis of the state variables. By assuming log normality, we then evaluate whether the probability of a funding ratio shortfall (i.e., a funding ratio smaller than one) in period t + 1 is less than δ. If this requirement is not met, those particular portfolio weights are excluded from the investment manager s choice set. As described above, the VaR can be imposed with respect to St (discounting at actual yields), S t (discounting at constant yields), or Ŝ t (discounting at the four-year average yield).

18 4 Results 4.1 The impact of smoothing yields In this section we investigate the investment manager s optimal portfolio choice when he is faced with risk constraints that are based on the smoothed liability measure. We consider a VaR constraint as the ex ante (preventive) constraint. For the ex post (punitive) constraint, we consider the requirement to draw AFCs whenever the plan becomes underfunded. Note again that both the VaR constraint and the AFCs are short-term considerations based on the smoothed liability measure whereas the long-term objective of wealth utility is defined with respect to the actual liability measure. We quantify in this section the welfare losses that result from the wedge that yield smoothing drives between these short- and long-term considerations. We first solve a one-period problem to explain the main intuition in a parsimonious setting. We then explain how the results change in a multi-period setup. Furthermore, it is very interesting to note that both the VaR constraint and AFCs are intended to decrease the manager s risky holdings as the funding ratio approaches the critical threshold of one. We show that when these short-term objectives are defined with respect to the smoothed liability measure, they can inadvertently induce the manager to increase his risky positions. We therefore conclude that smoothing yields may lead to highly perverse investment behavior and large welfare losses Case 1: ALM with a VaR constraint First we investigate optimal portfolio decisions and corresponding certainty equivalents in a one-period context (T = 1) under a VaR constraint. We set the state variables at time zero equal to their long-run averages. We set the VaR probability δ=0.025 and we do not include contributions from the sponsor (i.e., c T = 0). We compare a VaR constraint imposed on S T (discounting at actual yields) with one imposed on S T (discounting at a constant yield) and one on ŜT (discounting at the four-year average yield). Table 1 and 2 present the optimal portfolio weights and scaled certainty equivalents for constant and actual discounting for four different levels of risk aversion. Table 1 addresses the cases where risk aversion equals one and five, and in Table 2 we consider risk aversion levels of eight and 10. Note that the VaR constraint is more binding when the funding ratio at time zero, denoted by S 0, is lower. Therefore, as S 0 decreases, the manager has to substitute away from stocks to satisfy the constraint. The key insight of these results is that under actual discounting the manager substitutes into the long-term bond, whereas under constant discounting he moves into the riskfree asset. Because the utility from wealth

19 depends on the actual funding ratio, which is computed using current yields, investing in the riskless asset leads to large utility losses. The riskless asset does not hedge against liability risk and has a low expected return. In other words, when the VaR is imposed with constant discounting, the manager is torn between the objective of maximizing utility from wealth and satisfying the VaR constraint. When the VaR constraint is based on actual yields these two objectives are more aligned. The utility loss from constant discounting can be large and up to four percent of wealth. This loss is increasing in the degree of risk aversion. Substituting away from bonds into the riskless asset leaves a larger exposure to liability risk, leading to larger utility losses when the degree of risk aversion is higher. In other words, as risk aversion increases, the manager s preferred position in stocks is lower and he prefers to invest more in bonds to hedge against liability risk. As a consequence, the VaR constraint under actual discounting which requires a substantial weight in bonds does not affect the manager much. The VaR constraint under constant discounting, on the other hand, forces the manager into the riskfree asset leading to large welfare losses. A VaR constraint based on the smoothed liability measure may lead to a higher expected utility only when the degree of risk aversion is very low. This fact is easiest to understand in the following way. Suppose that the investment manager is risk neutral, meaning that he only cares about the average return on the portfolio. In steady state, the liabilities mainly add uncertainty to the problem which, in this case, the manager does not care about. Therefore, (ignoring the Jensen term) he maximizes the expected value of the assets only and can assume that the liabilities are constant. As a result, there is no longer a mismatch between the objective of maximizing utility from wealth and satisfying the VaR constraint. In fact, the VaR based on the smoothed liability measure may allow a higher weight in stocks, implying a higher expected return, and a higher certainty equivalent. The results above indicate that smoothing yields can lead to grossly suboptimal investment decisions. However, under current regulations, liabilities are discounted at the rolling four-year average yield, not at the unconditional average yield. As a consequence, the impact of smoothing is smaller than suggested above. However we show that it is still very large. When moving from period t to t + 1, the yield at time t 3 is dropped from the four-year average and the yield at time t + 1 is added. This implies that three values in the average stay the same and are known at time t. Only the yield at time t + 1 causes uncertainty. This has a relatively small impact on the average. In fact, as shown before, the conditional variance of the four-year average is a factor 16 smaller than the variance of the original series. Therefore, the investment manager still employs the riskfree asset to

20 satisfy the VaR constraint. To illustrate this effect in a one-period example, assume that in and before period zero the yields are steady state. In period one, reported liabilities are computed by discounting at 0.75ȳ y 15,t. Table 1 and 2 also present the optimal portfolio weights and scaled certainty equivalents for four-year average discounting, for risk aversion levels of one and five (Table 1) and eight and 10 (Table 2). As expected, the resulting portfolio weights are an average of those chosen under constant and actual discounting. The investment manager still wants to invest part of the funds in the riskfree asset. The welfare loss is still large and up to two percent of wealth. We conclude that imposing a VaR constraint on the smoothed funding ratio can lead to large welfare losses as it induces the use of the riskfree asset. Without the VaR constraint the investment manager would not use the riskfree asset because long-term bonds are a better hedge against long-run liability risk Case 2: ALM with AFCs We now assess the impact of smoothing yields by comparing optimal portfolio decisions and corresponding certainty equivalents when the investment manager has to request AFCs whenever the fund is underfunded. The notion of being underfunded depends strongly on the liability measure used. We show that AFCs based on the smoothed liability measure lead to a similar misalignment of objectives as under the VaR constraint. We set the contributions equal to the realized funding ratio shortfall, which implies c t = max(1 S t, 0) under constant discounting and c t = max(1 St, 0) under actual discounting. As before, we consider a oneperiod setup. We set λ > 1 to ensure concavity of the utility function and, for ease of exposition, we do not impose the VaR constraint. Finally, we set the state variables equal to their long-run averages at time zero. Table 3 presents the optimal portfolio weights and certainty equivalents for different values of λ and varying degrees of risk aversion as a function of the funding ratio S 0. The results show that when sponsor contributions and their consequent reputation loss are determined through constant discounting, the investment manager does not substitute into bonds but hedges against the utility penalties through the riskfree asset. This goes against the investment manager s desire to maximize wealth utility, leading to large welfare losses. There are, however, two main differences compared to the case of a VaR constraint. First, the manager now invests fully in stocks when the fund is highly underfunded, leading to a V- shaped policy function as in Berkelaar and Kouwenberg (2003). This is simply a consequence of our utility specification, which exhibits risk neutrality in the lower tail. Secondly, under constant discounting, as the funding ratio at time zero approaches the critical threshold of